Inferring molecular bonds from atomic coordinates is necessary when a reliable molecular graph is unavailable. We study the contribution of explicit hydrogens to this task using BondNet, a PaiNN-based network that jointly predicts connectivity and heavy–heavy bond types. On 27,240 GEOM-DRUGS test records, exact heavy–heavy graph recovery is 99.83% for clean coordinates and 69.56% at Gaussian coordinate noise $$\sigma =0.20$$ σ = 0.20 Å. The corresponding rates on a separately selected cohort of 10,000 PubChem3D molecules are 98.43% and 64.60%. At the higher noise level, default RDKit recovers 3.64% of PubChem3D graphs exactly; a Hückel configuration examined after the initial evaluation recovers 46.73%. A heavy-only BondNet reaches 50.70% exact match on GEOM at $$\sigma =0.20$$ σ = 0.20 , and the explicit-H advantage persists under matched heavy-atom perturbations. An oracle model supplied with reference hydrogen counts reaches 82.03%, suggesting that hydrogen-derived valence information contributes substantially to the observed benefit. Removing H atoms from the explicit-H model at inference sharply reduces bond-type accuracy. These results highlight the difference between bond-level and complete-graph accuracy. The GEOM test set was inspected during development, and the findings do not extend to charged molecules or authentic generated geometries. Scientific contribution We compare explicit-H and heavy-only input under the same training schedule and use an oracle hydrogen-count control to probe the information H atoms contribute. Heavy–heavy graph metrics, a separate PubChem3D cohort, and chemically normalized baselines show where learned bond perception succeeds and where high bond-level accuracy obscures reconstruction failures.
Pretrained chemical language models are selected, embedded with, and fine tuned on the basis of what their documentation says they are. We report that the released model card for MoLFormer, one of the most widely used chemical foundation models, states that isomeric information was removed from its training data, and that this statement is contradicted by artifacts distributed in the model's own repository. The shipped token frequency table records the tetrahedral stereocentre token [C@H] occurring 669,642,207 times, making it the twelfth most frequent token in a corpus of 47.2 billion tokens, ahead of fluorine, and stereochemical tokens account for 3.17
Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) prediction plays a central role in modern computer-aided drug discovery by enabling early evaluation of pharmacokinetic behaviour and safety liabilities. This tutorial presents a reproducible computational workflow for large-scale ADMET profiling using ADMET-AI 2.0, a graph neural network-based prediction platform, applied to 200 cyclin-dependent kinase 2 (CDK2) inhibitors curated from ChEMBL release 35. Implemented within an annotated Jupyter notebook, the workflow integrates cheminformatics preprocessing, physicochemical property analysis, drug-likeness evaluation, structural alert screening, absorption and distribution profiling, CYP450-mediated metabolism prediction, excretion assessment, toxicity evaluation, Tox21 pathway analysis, DrugBank percentile benchmarking, and composite multi-parameter optimisation scoring. A total of 104 ADMET-related outputs were generated per compound, including 52 predictive endpoints and 52 corresponding DrugBank approved-drug percentile rankings. The 200 curated CDK2 inhibitors exhibited 100
Drug discovery remains a costly and time-consuming process, motivating the development of artificial intelligence approaches to accelerate molecular design. Many generative models, including generative pre-trained transformers (GPT), rely on one-dimensional SMILES representations that lack explicit structural topology and are constrained by rigid syntax rules. Moreover, most existing models perform unconditional molecule generation, while the few conditional diffusion approaches often rely on abstract parameters rather than pharmacologically interpretable constraints. Consequently, current methods have limited ability to enforce explicit drug-like and pharmacophore requirements during molecular generation. Here, we propose GCMDiff, a multi-condition guided diffusion framework for molecular graph generation that supports controllable and interpretable molecular design. Unlike SMILES-based models, GCMDiff employs an atom-pair and bond-centric molecular graph representation that directly captures explicit chemical topologies. Crucially, we introduce an interpretable pharmacological condition embedding that integrates multi-modal constraints, including atom number, Lipinski’s Rule of Five constraints, and structural moieties derived from Checkmol, PubChem, and Rings in Drugs descriptors. These explicit features guide the diffusion process toward promising chemical spaces satisfying user-defined prerequisites. Benchmark experiments demonstrate that GCMDiff generates chemically valid and structurally diverse molecules, achieving 99.6 GCMDiff advances molecular design by employing a graph-based diffusion framework that bypasses SMILES-related syntax limitations. Its core innovation is a multi-modal condition encoder integrating drug-like properties and pharmacophores, enabling interpretable, pharmacophore-fragment-conditioned molecular generation. The model achieves 99.6
Public tumor-homing peptide (THP) resources have enabled predictors such as THPep, SCMTHP, StackTHPred, StackTHP and LLM4THP. At the same time, broader peptide machine-learning resources, including the recent PepBenchmark/PepBenchData initiative, are moving the field toward standardized datasets, preprocessing, and leaderboards across peptide-property tasks. These developments have improved comparability but have not removed task-specific problems in THP modelling: available THP benchmarks are motif-rich, chemically redundant, and susceptible to optimistic evaluation under near-neighbor overlap and potential leakage. A cheminformatics evaluation of THP predictions therefore needs to ask not only whether a peptide receives a high score, but also whether that score is independent, motif-driven, chemically redundant, hard-negative-sensitive, or near-duplicate-dependent. We present THP-NanoTarget, a reproducible peptide-cheminformatics framework for leakage-aware THP benchmarking, motif-rule comparisons, Morgan/ECFP chemical-space auditing, hard-negative evaluations, strict external-recognition filtering, and conservative analog prioritization. The primary benchmark contained 4202 peptides: 1401 THP positives and 2801 primary negatives. Under standard stratified out-of-fold evaluations, a term frequency-inverse document frequency (TF-IDF) 2–4-mer logistic-regression model achieved an area under the receiver operating characteristic curve (AUROC) of 0.784 and an area under the precision recall curve (AUPRC) of 0.716, improving over a transparent motif-rule baseline (AUROC 0.644; AUPRC 0.488). Cluster-aware validation was more conservative, with an AUROC of 0.708 and an AUPRC of 0.597 for the reference k-mer model and an AUROC of 0.715 and AUPRC of 0.612 for the best rank-averaged ensemble. Chemical-neighbor exclusion revealed local-neighborhood dependence: the AUROC declined to 0.721 at a Tanimoto of < 0.80 and 0.623 at < 0.70, while the 0.60 threshold was not evaluable. A benchmark redundancy analysis showed a median nearest-neighbor Morgan/extended connectivity fingerprint (ECFP) Tanimoto similarity of 0.837, and 95.8
